TL;DR
This paper introduces three web services designed to make the LiLa Knowledge Base of interoperable Latin linguistic resources more accessible. These services facilitate querying existing data and populating the knowledge base with new textual resources, aiming to bridge the gap between computational linguistics and classical studies.
What Problem Does This Paper Address?
The field of Linguistic Linked Open Data (LLOD) has made significant progress in developing vocabularies and ontologies to represent linguistic data and publishing these resources as Linked Open Data (LOD). However, a key challenge remains: making this interoperable data easily accessible and exploitable by a diverse range of users, including those without deep expertise in semantic web technologies like SPARQL. This is particularly crucial for fields like Classics, where computational tools and skills are not always widespread. The LiLa Knowledge Base, a project aiming to interlink Latin linguistic resources as LOD, faces this challenge directly. The paper addresses this by presenting services that simplify querying and contributing to the LiLa KB, thereby lowering the barrier to entry for classicists and computational linguists alike.
Methods
The paper describes three core services, all developed as web applications with a React JavaScript frontend and servlet-based backend:
- LiLa Knowledge Base (KB): This forms the foundation, integrating various linguistic resources for Latin using established ontologies (OLIA, Ontolex-Lemon, POWLA) and Semantic Web standards (RDF). Its architecture is lexically-based, using lemmas as the primary interface.
- Lemma Bank Query Interface: This service allows users to search the LiLa KB's collection of Latin lemmas. Users can filter by lemma string (supporting regular expressions), prefixes, suffixes, lexical bases, gender, part of speech (PoS), and inflectional category. A key feature is its dynamic query decomposition, which optimizes results by recalculating filter options as the user refines their search.
- LiLa Interactive Search Platform (LISP): This graphical interface enables SPARQL queries on both textual and lexical resources within the LiLa RDF triple store. LISP visualizes the relationships between resources as a network, allowing users to build queries by selecting nodes (e.g., Authors, Documents, Tokens, Lemma Bank, Lexical Resources). It employs a depth-first search algorithm to construct SPARQL queries dynamically.
- LiLa Text Linker: This web application assists in producing RDF editions of Latin texts. It automates tokenization, lemmatization, and PoS tagging using a custom UDPipe model trained on a substantial corpus of Latin data. It then matches these tokens against the LiLa KB and allows for manual correction of matches. The output can be exported as RDF.
Datasets
The paper does not introduce new benchmark datasets. Instead, it leverages and integrates existing linguistic resources for Latin, including:
- LEMLAT: A morphological analyzer database.
- Opera Latina corpus: A collection of 130 Classical Latin texts.
- Index Thomisticus Treebank: Texts by Thomas Aquinas.
- UDante treebank: Medieval Latin works by Dante Alighieri.
- Lewis and Short dictionary: A primary resource for Classical Latin.
- Dictionary of Medieval Latin in the Czech Lands: A lexical resource for Medieval Latin.
- Word Formation Latin: A derivational lexicon.
- Latin WordNet: Enhanced with valency information.
- LatinAffectus polarity lexicon.
- Latin Vallex lexicon.
The LiLa Text Linker's ad-hoc model for tokenization, lemmatization, and PoS tagging was trained on approximately 3.4 million tokens from various Latin treebanks (Index Thomisticus, PROIEL, Perseus, UDante), the Opera Latina corpus, the Computational Historical Semantics database, and other curated lemmatized works.
Key Insights and Ideas
- Bridging the Humanist-Computational Divide: The core insight is that making complex linguistic data and tools accessible through user-friendly interfaces is paramount for fostering interdisciplinary collaboration. The LiLa services are designed to empower classicists with computational methods without requiring them to become experts in semantic web technologies.
- Lemma as the Central Hub: The LiLa KB's architecture, centered around lemmas, proves to be a highly effective strategy for interoperability. Lemmas act as a common ground to link diverse lexical and textual resources.
- Interactive Querying for Exploration: The graphical interfaces of the Lemma Bank Query Interface and LISP are crucial for making complex querying intuitive. Visualizing relationships and dynamically updating filter options significantly lowers the cognitive load for users.
- Automated Text Processing for RDF Creation: The LiLa Text Linker streamlines the laborious process of converting raw texts into structured RDF data, a critical step for integrating new resources into the LOD ecosystem.
Key Results
- The LiLa KB contains over 80 million triples, integrating numerous lexical and textual resources for Latin.
- The LiLa Text Linker's ad-hoc UDPipe model achieved high performance for lemmatization and PoS tagging, with precision and recall around 94% for UPOS and 93.7% for lemmas.
- The services are presented as functional web applications, demonstrating the practical implementation of interoperable linguistic data for Latin.
Limitations and Future Work
- Portability of Tools: While the services are designed for LiLa, their adaptability to other LLOD projects using different ontologies or data structures is a key area for future work. The Text Linker currently relies on an SQL database for efficiency, but adapting it to interface with triple stores is planned to enhance portability.
- Usability Testing: The authors plan to conduct more extensive usability tests with representatives from different user communities to further refine the interfaces.
- Expanding Resource Access: LISP will be enhanced to include access to more lexical resources, such as the "Lexikon der indogermanischen Verben," by generalizing its query process.
- Improving Text Linker: The Text Linker's trained model will be updated using a larger training set and a newer version of UDPipe. Support for additional formats like TEI-compliant XML is also planned.
- User Adoption: The ultimate goal is for classicists to not only use these services but also to become more autonomous in querying and publishing linguistic data themselves.
Summary updated